Visual analytics methods for supporting semantic abstraction of multivariate time series data

Abstract

This dissertation focuses on the challenges and methods in analyzing multivariate time series (MVTS) data, crucial in various fields like finance, healthcare, and sports. The raw MVTS data, often complex due to its volume and multiple attributes, limits insightful pattern recognition. Temporal abstraction is proposed as a solution, integrating elementary data (i.e., sequences of time-stamped attribute values) into meaningful entities for easier comprehension and analysis. At the basic level, the abstraction transforms data referring to time points into interval-based representations. At higher levels, temporal relations between earlier derived representations are used to integrate them into more complex concepts. <br /> Previous research in MVTS analysis has two main directions: computational/algorithmic methods and visual analytics. The former focuses on extracting patterns from data but is not concerned with enabling human analysts to consider the contexts in which the patterns occur and the relationships between the patterns. Visual analytics research encompasses several application-specific studies but lacks a systematic approach to incorporating temporal abstraction in analytical workflows. This dissertation aims to bridge these gaps by developing a framework for integrating computational methods with techniques for interactive visual analysis designed to support human cognition and knowledge generation. <br /> The thesis focuses on the extraction, interpretation, and analysis of patterns from MVTS through progressive abstraction. It starts with methods for detecting temporal patterns in individual variables, progresses to deriving higher-level patterns by combining univariate patterns, and works out approaches to exploring the distribution of these patterns across the dataset. In all these steps, it addresses the problem of supporting human understanding and analytical reasoning by means of effective visualizations.

The dissertation includes the following major parts: <br /> First, it focuses on detecting patterns such as up-trends and peaks in discretized MVTS, employing geometric rules and visualization for pattern recognition. <br /> Second, it explores identifying patterns in continuous MVTS, using computational algorithms to understand patterns and their temporal relations. <br /> Lastly, it introduces topic modeling to examine concurrency between multiple univariate patterns in discretized MVTS. <br /> In conclusion, this dissertation provides a conceptual framework and methods for progressive temporal abstraction in MVTS data. It advances the field by combining computational and visual techniques to aid domain experts in data interpretation. Future research involves adapting and refining this framework across different domains, enhancing temporal pattern analysis, and incorporating expert feedback for continual improvement and broader applicability.